R600/RegisterCoalescer: Enable more rematerialization/add missing testcase

This enables the rematerialization of some R600 MOV instructions in the
RegisterCoalescer and adds a testcase for r235668.

llvm-svn: 235675

diff --git a/llvm/lib/Target/R600/SIInstructions.td b/llvm/lib/Target/R600/SIInstructions.td
index dbd24d3..bac9274 100644 (file)
--- a/llvm/lib/Target/R600/SIInstructions.td
+++ b/llvm/lib/Target/R600/SIInstructions.td
@@ -97,7 +97,7 @@ defm S_BUFFER_LOAD_DWORDX16 : SMRD_Helper <
 //===----------------------------------------------------------------------===//
 
 let isMoveImm = 1 in {
-  let isReMaterializable = 1 in {
+  let isReMaterializable = 1, isAsCheapAsAMove = 1 in {
     defm S_MOV_B32 : SOP1_32 <sop1<0x03, 0x00>, "s_mov_b32", []>;
     defm S_MOV_B64 : SOP1_64 <sop1<0x04, 0x01>, "s_mov_b64", []>;
   } // let isRematerializeable = 1
@@ -1234,7 +1234,7 @@ let vdst = 0, src0 = 0 in {
 defm V_NOP : VOP1_m <vop1<0x0>, (outs), (ins), "v_nop", [], "v_nop">;
 }
 
-let isMoveImm = 1 in {
+let isMoveImm = 1, isReMaterializable = 1, isAsCheapAsAMove = 1 in {
 defm V_MOV_B32 : VOP1Inst <vop1<0x1>, "v_mov_b32", VOP_I32_I32>;
 } // End isMoveImm = 1
 

diff --git a/llvm/test/CodeGen/R600/coalescer_remat.ll b/llvm/test/CodeGen/R600/coalescer_remat.ll
new file mode 100644 (file)
index 0000000..f78a77b
--- /dev/null
+++ b/llvm/test/CodeGen/R600/coalescer_remat.ll
@@ -0,0 +1,57 @@
+; RUN: llc -march=amdgcn -verify-machineinstrs -o - %s | FileCheck %s
+target triple="amdgcn--"
+
+declare float @llvm.fma.f32(float, float, float)
+
+; This checks that rematerialization support of the coalescer does not
+; unnecessarily widen the register class. Without those fixes > 20 VGprs
+; are used here
+; Also check that some rematerialization of the 0 constant happened.
+; CHECK-LABEL: foobar
+; CHECK:  v_mov_b32_e32 v{{[0-9]+}}, 0
+; CHECK:  v_mov_b32_e32 v{{[0-9]+}}, 0
+; CHECK:  v_mov_b32_e32 v{{[0-9]+}}, 0
+; CHECK:  v_mov_b32_e32 v{{[0-9]+}}, 0
+; CHECK: ; NumVgprs: 12
+define void @foobar(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in, i32 %flag) {
+entry:
+  %cmpflag = icmp eq i32 %flag, 1
+  br i1 %cmpflag, label %loop, label %exit
+
+loop:
+  %c = phi i32 [0, %entry], [%cnext, %loop]
+  %v0 = phi float [0.0, %entry], [%fma.0, %loop]
+  %v1 = phi float [0.0, %entry], [%fma.1, %loop]
+  %v2 = phi float [0.0, %entry], [%fma.2, %loop]
+  %v3 = phi float [0.0, %entry], [%fma.3, %loop]
+
+  ; Try to get the 0 constant to get coalesced into a wide register
+  %blup = insertelement <4 x float> undef, float %v0, i32 0
+  store <4 x float> %blup, <4 x float> addrspace(1)* %out
+
+  %load = load <4 x float>, <4 x float> addrspace(1)* %in
+  %load.0 = extractelement <4 x float> %load, i32 0
+  %load.1 = extractelement <4 x float> %load, i32 1
+  %load.2 = extractelement <4 x float> %load, i32 2
+  %load.3 = extractelement <4 x float> %load, i32 3
+  %fma.0 = call float @llvm.fma.f32(float %v0, float %load.0, float %v0)
+  %fma.1 = call float @llvm.fma.f32(float %v1, float %load.1, float %v1)
+  %fma.2 = call float @llvm.fma.f32(float %v2, float %load.2, float %v2)
+  %fma.3 = call float @llvm.fma.f32(float %v3, float %load.3, float %v3)
+
+  %cnext = add nsw i32 %c, 1
+  %cmp = icmp eq i32 %cnext, 42
+  br i1 %cmp, label %exit, label %loop
+
+exit:
+  %ev0 = phi float [0.0, %entry], [%fma.0, %loop]
+  %ev1 = phi float [0.0, %entry], [%fma.1, %loop]
+  %ev2 = phi float [0.0, %entry], [%fma.2, %loop]
+  %ev3 = phi float [0.0, %entry], [%fma.3, %loop]
+  %dst.0 = insertelement <4 x float> undef,  float %ev0, i32 0
+  %dst.1 = insertelement <4 x float> %dst.0, float %ev1, i32 1
+  %dst.2 = insertelement <4 x float> %dst.1, float %ev2, i32 2
+  %dst.3 = insertelement <4 x float> %dst.2, float %ev3, i32 3
+  store <4 x float> %dst.3, <4 x float> addrspace(1)* %out
+  ret void
+}